Beeler Patrick, Dalleur Olivia, Bates David

Introduction

Having structured diagnoses is crucial for inpatient treatment and has specific implications regarding clinical decision support (CDS) and billing. Nevertheless, most diagnoses are not electronically available until the coding staff review the patient charts, i.e., after discharge. This lack of structured data on diagnoses of hospitalised patients may affect the quality of care. However, electronic health records provide current, structured data on treatment orders and drug agents have been shown to be associated with medical conditions in the outpatient setting. As yet, little is known about associations of administration routes and doses with diagnoses, especially in inpatients. We compared the prediction of diagnoses using (i) drug orders alone vs drugs and routes, and (ii) drugs and routes vs drugs, routes and doses.

Methods

The study was a retrospective cohort study, conducted at the Brigham and Women’s Hospital, Boston, MA. Inpatients, hospitalised in any ward, were included if they had been discharged between 1 Jan. 2012 and 31 Dec. 2012. All diagnoses coded according to the International Classification of Diseases version 9 (ICD-9) and data on all drugs related to these hospitalisations were obtained from the clinical information system in order to build drug-diagnosis pairs and then include stepwise route and dose. The dose was categorised into high (≥mean) and low (<mean). Only pairs present in ≥5 hospitalisations were analysed (‘support’). In order to compute the sensitivity and specificity of each association, hospital stays with the investigated drug-diagnosis pair (true positives), the drug only (false positives), only the diagnosis (false negatives), and with neither the drug nor the diagnosis (true negatives) were counted. These numbers were computed, first for the drug-diagnosis pairs, second for the drug-and-route-diagnosis pairs, and third for the drug-with-route-and-dose-diagnosis pairs. Theχ² statistic was also calculated for each pair to indicate the corresponding prediction strength. Finally, the associations were juxtaposed, comparing (i) drug orders alone with drugs and routes, and (ii) drugs and routes with drugs, routes and doses. The difference between the χ² statistics of the juxtaposed pairs allowed for the comparison and presentation of the improved prediction after including routes and doses.

Table 1 Top 10 out of 310 102 pairs: comparison of the sensitivity and specificity of drugs predicting diagnoses with drugs including route information.
Drug agent	Sensitivity (%)	Specificity (%)	By including this route	Sensitivity (%)	Specificity [%]	Predicted diagnosis	χ2 statistic gain by including route
Vancomycin	82.5	80.9	PO	69.5	99.6	Intestinal infection due to Clostridium difficile	21 920.2
Acetylcysteine	77.8	98.4	IV	72.2	99.9	Poisoning by aromatic analgesics, not elsewhere classified	10 587.0
Sodium chloride	73.1	89.2	INH	36.8	99.8	Cystic fibrosis with pulmonary manifestations	9 100.0
Epinephrine	54.6	98.4	IM	51.5	99.8	Need for desensitisation to allergens	8 497.0
Cromoglicic acid	54.5	100.0	PO	54.5	100.0	Congenital pigmentary anomalies of skin	7 514.5
Rufinamide	33.3	100.0	GTUBE or JTUBE	33.3	100.0	Congenital quadriplegia	7 348.4
Nalbuphine	41.7	95.2	IV	37.6	99.3	Outcome of delivery, single liveborn	6 607.1
Paracetamol	93.7	27.2	PR	50.8	95.7	Acute posthaemorrhagic anaemia	6 490.2
Ondansetron	93.9	51.9	PO	84.6	93.5	Encounter for antineoplastic chemotherapy	6 459.0
Insulin (human)	65.3	78.5	IV	44.7	97.4	Acute posthaemorrhagic anaemia	6 194.3

Results

A total of 55 095 hospital stays were analysed. The most common diagnoses were ‘unspecified essential hypertension’ (coded for 29.5% of stays), ‘other and unspecified hyperlipidaemia’ (16.9%), ‘oesophageal reflux’ (15.6%), ‘outcome of delivery, single liveborn’ (13.1%), and ‘personal history of tobacco use’ (12.6%). Most patients received pneumococcal vaccines (used in 74.8% hospitalisations), paracetamol (73.4%), docusate sodium (63.6%), oxycodone (52.8%) and ondansetron (48.8%). Overall, 745 distinct drug agents, 27 administration routes, 2 dose levels and 3 616 diagnoses formed prediction pairs. The comparison of the χ² statistics of drugs predicting diagnoses showed substantial improvements after including the route (table 1). For example, the specificity of vancomycin predicting ‘intestinal infection due to Clostridium difficile’ increased 18.7% by including the route ‘peros’, though the sensitivity decreased by 13%. The most substantial χ² gain due to further inclusion of dose information was calculated for high-dose diphenhydramine per os, associated with ‘outcome of delivery, single liveborn’, increasing the specificity by 21.6% and decreasing the sensitivity by only 0.4% (table 2).

Table 2 Top 10 out of 351 266 pairs: comparison of the sensitivity and specificity of drugs and routes predicting diagnoses with drugs including route and dose information.
Drug agent	Route	Sensitivity (%)	Specificity (%)	Dose	Unit	Sensitivity (%)	Specificity (%)	Predicted diagnosis	χ2 statistic gain by including dose
Diphenhydramine	PO	92.6	76.7	High	mg	92.2	98.3	Outcome of delivery, single liveborn	30 120.2
Multivitamins1	PO	82.1	65.4	High	Tab	62.8	99.9	Cystic fibrosis with pulmonary manifestations	21 686.6
Coagulation factor IX	IV	71.4	99.9	High	Unit	71.4	100.0	Congenital factor IX disorder	14 500.5
Oxycodone	PO	88.5	53.1	High	mg	84.5	84.1	Outcome of delivery, single liveborn	11 408.4
Diphenhydramine	PO	91.8	70.9	High	mg	91.2	90.6	Other current conditions of mother, delivered3	10 455.2
Calcium folinate	PO	47.2	99.4	High	mg	47.2	100.0	Malignant neoplasm of long bones of lower limb	9 969.4
Ethambutol	PO	83.3	99.9	High	mg	83.3	100.0	Pulmonary alveolar proteinosis	9 663.8
Multivitamins1	PO	70.2	65.4	High	Tab	47.8	99.7	Bronchiectasis with acute exacerbation	9 042.6
Pseudoephedrine	PO	61.3	99.6	High	mg	60.7	99.9	Benign neoplasm of pituitary gland4	8 721.8
Multienzymes2	PO	87.9	99.2	High	Tab	74.9	99.8	Cystic fibrosis with pulmonary manifestations	8 005.8
1 ‘Multivitamins, other combinations’ 2 ‘Multienzymes (lipase, protease, etc.)’ 3 ‘Other current conditions classifiable elsewhere of mother, delivered, with or without mention of antepartum condition’ 4 ‘Benign neoplasm of pituitary gland and craniopharyngeal duct’

Discussion

We found that consideration of not only the drug, but also the route and the dose, had a substantial impact on the ability to predict inpatient diagnoses in real time. This has important implications: knowledge of the diagnoses early in the process may directly improve the quality of care; in addition, the present approach allows for automated pick lists displaying the most likely diagnoses, has the potential to minimise misunderstandings between healthcare professionals, to enable sophisticated CDS and to support efforts aiming to increase the completeness of problem lists.

Disclosure statement

Dr. Beeler was supported by the Swiss National Science Foundation.

Correspondence

Correspondence:

Dr. med. Patrick E. Beeler

Research Center for Medical Informatics

University Hospital Zurich

Rämistrasse 100

CH-8091 Zurich

Switzerland

patrick.beeler[at]usz.ch

Copyright

Published under the copyright license
“Attribution – Non-Commercial – NoDerivatives 4.0”.
No commercial reuse without permission.
See: emh.ch/en/emh/rights-and-licences/